In this talk we will discuss our experience building AI systems for enterprise process automation. Using examples of real-life deployed AI systems in AdTech, customer service, mortgage financing and recruiting, we discuss our learnings and insights gleaned.
2. 22
AI-driven enterprise applications
•Business processes mapped to an AI engine to enable
business efficiencies.
•4 business processes being automated by AI
Customer Support
Recruiting
Content Marketing
AdTech
• We will conclude with lessons learned from being very
involved in these companies since inception.
3. 33
But then what is AI? – Lessons Learned
•AI is a rich source of interesting tools
Lot more than Deep Learning, CNN, Generative Adversarial
Networks!!
Suite of techniques to evoke intelligence :
Categorizers, Regression, NLP, Case-Based reasoning etc.
•Domain driven rather than technique driven
Let the domain drive the problem solving and which techniques
you use from the bag
•Interesting Data strategies
•AI application is like a raisin bread : it is still 90% bread
5. 55
Why Neva?
Customer service organizations must improve
support quality while reducing delivery costs.
Key challenges
Fragmented knowledge from disparate knowledge sources and
enterprise systems, and decentralized change management.
Inefficient decision-making due to gap between front and back
office, frequent changes, and inability to continuously train human
agents.
Fractured user experience due to omni-channel, modern support
outside work and inefficient, human-based support at work
10. www.swooptalent.com
Your PRIVATE data
backbone
Data from ALL sources
matched & made available
Private Talent Data Cloud
Production
ATS - cloud
Data from
prior ATS
CRM and
other live
systems -
cloud
Hundreds of millions of social talent records gathered by Swoop
Resumes,
Spreadsheets,
etc
12. www.swooptalent.com
Swoop AI Layer
More Structured Less Structured
ATS, CRM
XML, Excel, Flat
Files,
Social Media, Niche
Forums, Society Boards
Docs, PDF, JPG,
Supervised Learning
Tokenization
Part of
Speech
Named Entity Recognition
Custom Pipelines
Unsupervised Learning
Clustering Similarity
Latent Semantic Analysis,
SVD, Word2Vec
Topic Modeling (160 Million Profiles)
Data Data
Data Data
Search
Semantic
Query
Processing
Topic
Modeling
Application
15. Enterprises Journey to Autonomous Marketing
Data
Sync
Banks Data
Social Data
Public Records
Data
Cleansing
Data is Engineered
Content
Creation
Search Content
Social Content
Email & Text
Machine
Learning
NLPK
Data Science
• Markovian Modeling
16. Up IQ to Power Banks: SEM Campaigns, & Landing Pages
17. Customers Journey, from Discovery to Acquisition
Personalized
Banks
Retail Banks
Mortgage Banks
Online Lenders
Relevant
Bank Staff
Ranked Bank Staff
Content
Discovery
Search Content
Social Content
Email & Text
• Information Theoretic Scoring
• Sentiment Analysis
19. Cross Device Graph
1919
Machine learning models device graph
relationships : naive Bayes modeling &
heuristics for pruning.
172.0.0.217
Feature engineering (UID, IP, user agent,
referral url, login email etc.)
Data Collection (cookie-sync, exchanges,
ad impression, native sdk, 3rd party data
Identify users across smartphones, tablets & desktops
20. Bid Price Optimization
20
• A dynamic pricing algorithm
– maximizes the expected value of gain after winning an auction, or 𝑏 =
𝑎𝑟𝑔𝑚𝑎𝑥 𝑏 𝐸 𝑔𝑎𝑖𝑛
– adjusts automatically to meet business requirements (ex. CPM margin)
using a feedback loop
auction data
user data
win rate
win price
purchase
prediction
ctr
bidding strategy
bid price
business
requirements
alpha
• Machine learning models
win rate – binary classification (Random forest)
win price – regression
purchase prediction – binary classification
(Random forest)
CTR – binary classification
21. 2121
But then what is AI? – Lessons Learned
•AI is a rich source of tools
Deep Learning, CNN, Generative Adversarial Networks
Categorizers, Regression, NLP, Case-Based reasoning etc.
•Domain driven rather than technique driven
•Interesting Data strategies
•AI application is like a raisin bread : it is still 90% bread